74.5%
▲ 44.2 pts since Jan 2025
GAIA accuracy of Claude Sonnet 4.5, September 2025; human baseline 92%
20252025
What the chart shows
GAIA tests general AI assistants on multistep, real-world questions that typically need web browsing, file handling and reasoning across sources. The best score rose from 30.3% for DeepSeek R1 (January 2025) to 74.5% for Claude Sonnet 4.5 (September 2025).
In between, Claude 3.7 Sonnet (February) reached 64.2%, Claude Opus 4 (May) 64.8% and Claude Opus 4.1 (August) 68.5%. The human baseline is 92%, so the leading model remains 17.5 percentage points behind.
Key numbers
| Measure | Value |
|---|---|
| DeepSeek R1, Jan 2025 | 30.3% |
| Claude 3.7 Sonnet (high), Feb 2025 | 64.2% |
| Claude Opus 4 (high), May 2025 | 64.8% |
| Claude Opus 4.1 (high), Aug 2025 | 68.5% |
| Claude Sonnet 4.5, Sep 2025 | 74.5% |
| Human baseline | 92% |
How to read it
- Each bar is the best model released by that month on the GAIA leaderboard the AI Index cites (Princeton’s HAL); months with no new best are left out.
- The report text gives about 20% for January 2025; the data shows 30.3% (DeepSeek R1), which we use.
Source: GAIA Leaderboard, 2026, via the AI Index Report 2026 (Stanford HAI), figure 2.6.1: GAIA: accuracy. Chart and text by AI Adoption.
Numbers checked against the published data on 20 October 2026.
Download the data (CSV) · Cite this chart
Numbers checked against the published data on 20 October 2026.
Download the data (CSV) · Cite this chart
Discussion